News for hackers. Stuff that matters.
Re:Djbsort: A new software library for sorting arrays of int... by MrBuddyCasino on Wednesday July 11, 2018 @05:17AM (#17505524)
The 2.5 cycles/byte compared to the 32 cycles/byte that Intel managed pulled off seems like an improbably large improvement over the current state of the art? Is this real?
Re:Djbsort: A new software library for sorting arrays of int... by Cyphase on Wednesday July 11, 2018 @05:28AM (#17505561)
The cycles/byte for all the sizes listed in the table on the Speed page[0], courtesy of copy-paste and a one-liner in the Python REPL:

  size      cycles/byte (based on median)
  1         6.0
  2         3.375
  4         11.625
  8         9.625
  16        7.984375
  32        7.515625
  64        6.0390625
  128       4.31640625
  256       2.4873046875
  512       2.29443359375
  1024      2.5048828125
  2048      2.77893066406
  4096      3.0791015625
  8192      3.89566040039
  16384     5.23690795898
  32768     6.35472106934
  65536     7.55914306641
  131072    9.23189163208
  262144    10.789557457
  524288    12.4885950089
  1048576   14.6550707817
[0] https://sorting.cr.yp.to/speed.html
Re:Djbsort: A new software library for sorting arrays of int... by Ono-Sendai on Wednesday July 11, 2018 @05:20AM (#17505531)
Well, that's for a very small and particular size. (1024 elements) Intel's library might have more overhead at that size.
Re:Djbsort: A new software library for sorting arrays of int... by bdukic on Wednesday July 11, 2018 @05:22AM (#17505541)
My guess is that the "catch" comes in the form of the list of limitations.